Plurai, a startup founded by two former Nexar engineers, has raised $13 million in seed funding to fix one of the most stubborn problems in enterprise AI: agents that look impressive in a demo and fall apart in production. The round was led by Team8, with participation from U&I, Velocity and Mercer, and Google DeepMind researcher Reah Miyara joined as an angel investor.
The idea is borrowed from the world the founders came from. Co-founders Ilan Kadar and Elad Levi previously worked at Nexar, which builds autonomous driving technology, and Kadar worked with Waymo on simulating and evaluating its end-to-end deep learning driving software. Self-driving cars only became trustworthy once they could be tested against millions of simulated edge cases before touching a real road. Plurai wants to do the same thing for AI agents.
The demo problem
"The moment we recognized the surge of AI agents, we noticed a familiar trend: it's easy to create a demo, but [challenging] to develop something reliable and cost-effective," Kadar told TechCrunch. Like the end-to-end models trained to drive cars, AI agents are probabilistic systems operating in unpredictable environments, and the failures that matter most are the rare ones — the unusual request, the oddly formatted document, the tool that returns something unexpected.
That is the gap Plurai is trying to close. Rather than letting an agent learn from live mistakes on real company data, the startup builds a simulation layer first, so the messy cases can be found and fixed before anything goes wrong in production.
Digital twins for agents
Plurai creates a digital twin of a company's electronic systems, including the knowledge base and the tools an agent will have access to in production. It then runs every possible path an agent might take through those systems, using synthetic data to prepare models for difficult edge cases and to train on the spots where things break.
The platform is designed to work alongside existing agentic tools such as Claude Code, Codex, Copilot and Cursor, rather than replacing them. The simulation layer sits underneath, generating the training data that makes those tools behave more predictably when they are pointed at a specific company's systems.
Small models, frontier results
The company's longer-term pitch is to use the data gathered in simulation to post-train small, open-weight models tailored to individual customer needs. In tests performed in cooperation with Nvidia and Nebius, Plurai says its post-trained models matched or beat mid-tier frontier models at sharply lower cost — one to two orders of magnitude cheaper, according to the company.
If that holds up, it points to a different economics for enterprise agents: instead of renting the biggest model for every task, companies could run smaller, cheaper models that have been trained specifically on their own workflows and failure modes. Plurai is entering a crowded field — agents are one of the most funded corners of AI — but its bet is that reliability, not raw capability, is what will decide which agents survive in production.



